mirror of
https://github.com/libretro/virtualjaguar-libretro.git
synced 2026-11-12 16:21:34 -10:00
Joseph MattielloandClaude Sonnet 5
b4d0c999c8
The GPU entry point re-wrote MB_SENT_START/MB_MAGIC_START, duplicating what the 68K bootstrap already writes at cart boot per Task 1's cartboot.inc convention. Provably inert (same value, before the timed loop) but dead and confusing -- remove it so Task 3 doesn't copy the pattern forward. Regenerated benchgpu_arith.j64 via build.sh; count still reads 2000000 (done_frame=86, unchanged) via make microbench. Co-Authored-By: Claude Sonnet 5 <[email protected]>
112 lines
4.8 KiB
ArmAsm
112 lines
4.8 KiB
ArmAsm
; test/microbench/benchgpu_arith_gpu.s -- GPU-side code for the GPU
|
|
; arithmetic-loop microbenchmark (#536, task 2).
|
|
;
|
|
; Assembled separately with lyxass (the GPU RISC has its own ISA, not the
|
|
; 68000's), then the 12-byte BS94 header is stripped and the raw code is
|
|
; embedded into the 68K bootstrap (benchgpu_arith.s) via .incbin. See
|
|
; README.md "For tasks 2-5: embedding a lyxass GPU/DSP blob".
|
|
;
|
|
; Assemble:
|
|
; lyxass -o benchgpu_arith_gpu.o benchgpu_arith_gpu.s
|
|
; tail -c +13 benchgpu_arith_gpu.o > benchgpu_arith_gpu.bin ; strip BS94 header
|
|
;
|
|
; The 68K loader copies this blob to GPU local RAM at $F03000 and starts
|
|
; execution there, so `.org $F03000` here matches where it will actually
|
|
; run (register-relative code, but jr targets are PC-relative so this
|
|
; mostly just documents intent).
|
|
;
|
|
; --- workload shape -----------------------------------------------------
|
|
; Per #536's dispatch-vs-body framing: this ROM isolates GPU *arithmetic*
|
|
; throughput -- the loop body is almost entirely ALU ops (ADD/SUB/AND/OR
|
|
; class), with exactly one taken branch (JR) closing the loop. Task 3
|
|
; (benchgpu_branch) is the mirror image: branch-heavy, ALU-light.
|
|
;
|
|
; Body per iteration: 16 ALU ops (add/sub/and/or x4) + ADDQ (count++) +
|
|
; SUBQ (countdown--) = 18 arithmetic instructions, then a conditional
|
|
; JUMP (Rn) + its mandatory delay-slot NOP = 2 control instructions.
|
|
; 18/20 = 90% arithmetic. JUMP rather than JR: the loop body is 20
|
|
; instructions long and JR's signed relative offset can't span that (see
|
|
; the `movei #loop,r11` comment below -- measured, not assumed).
|
|
;
|
|
; Registers r3-r10 are scratch accumulators fed into each other so the
|
|
; assembler/optimizer (there isn't one here, but for clarity) can't elide
|
|
; the work; their final values are never read back -- only the iteration
|
|
; count in r1 and the fixed magics matter to the harness.
|
|
|
|
MB_SENT_BASE .equ $00010000
|
|
MB_SENT_START .equ MB_SENT_BASE+0
|
|
MB_SENT_DONE .equ MB_SENT_BASE+4
|
|
MB_SENT_COUNT .equ MB_SENT_BASE+8
|
|
|
|
MB_MAGIC_START .equ $C0DE57A7
|
|
MB_MAGIC_GPUAR .equ $C0DE0A01
|
|
|
|
; ITERATIONS: 2,000,000 passes of the 20-instruction body above. Nominal
|
|
; cost per pass (1 cycle/ALU op, ignoring load/store and branch-taken
|
|
; pipeline effects the JTRM documents but which don't apply here since
|
|
; there are no LOAD/STORE in the hot loop): ~20 cycles, so ~40M cycles at
|
|
; the GPU's ~26.6 MHz full-system clock ~= 1.5s of emulated time. See
|
|
; README.md for the measured completion frame -- re-measure and update
|
|
; both this comment and the tool's budget block if this value changes.
|
|
ITERATIONS .equ 2000000
|
|
|
|
.gpu
|
|
.org $F03000
|
|
.run
|
|
start:
|
|
; MB_SENT_START/MB_MAGIC_START are written by the 68K at cart boot
|
|
; (cartboot.inc convention, Task 1) -- do not write them again here.
|
|
moveq #0,r1 ; iterations completed
|
|
movei #ITERATIONS,r2 ; countdown
|
|
|
|
; seed the scratch accumulators so the ALU ops have real operands
|
|
moveq #1,r3
|
|
moveq #2,r4
|
|
moveq #3,r5
|
|
moveq #4,r6
|
|
moveq #5,r7
|
|
moveq #6,r8
|
|
moveq #7,r9
|
|
moveq #8,r10
|
|
movei #loop,r11 ; JR's signed offset can't reach
|
|
; across a 20-instruction body --
|
|
; measured: lyxass rejects it
|
|
; ("Relative distance to large").
|
|
; JUMP (Rn) has no such range limit.
|
|
|
|
loop:
|
|
add r3,r4 ; 1
|
|
sub r5,r6 ; 2
|
|
and r7,r8 ; 3
|
|
or r9,r10 ; 4
|
|
add r4,r3 ; 5
|
|
sub r6,r5 ; 6
|
|
and r8,r7 ; 7
|
|
or r10,r9 ; 8
|
|
add r3,r4 ; 9
|
|
sub r5,r6 ; 10
|
|
and r7,r8 ; 11
|
|
or r9,r10 ; 12
|
|
add r4,r3 ; 13
|
|
sub r6,r5 ; 14
|
|
and r8,r7 ; 15
|
|
or r10,r9 ; 16
|
|
addq #1,r1 ; 17 (count++)
|
|
subq #1,r2 ; 18 (countdown--)
|
|
jump ne,(r11) ; 19 (control)
|
|
nop ; 20 (delay slot -- control overhead)
|
|
|
|
; Count first, magic second: a reader that sees DONE is then
|
|
; guaranteed to see a settled count as well (same ordering rule
|
|
; as bench68k.s / cartboot.inc).
|
|
movei #MB_SENT_COUNT,r14
|
|
store r1,(r14)
|
|
movei #MB_MAGIC_GPUAR,r0
|
|
movei #MB_SENT_DONE,r14
|
|
store r0,(r14)
|
|
|
|
halt:
|
|
jr halt
|
|
nop
|
|
.end
|